AI Model Management

[mmccool]

Session description

AI models can be executed on the client web platform and can add significant functionality to web applications. However, they can also be quite large, requiring significant resources to download and store. Download and compilation latencies can potentially impact the user experience.

This breakout will discuss ways in which these issues can be mitigated. Possible topics include the following.

• Background model download and compilation.
• Caching strategies, including potential cross-site caching mechanisms with privacy-preserving mitigations
• Model naming and versioning, allowing for model substitution when useful
• Access to both downloadable and pre-installed models with a common interface
• Storage deduplication
• Model representation independence
• API independence (e.g. sharing models between WebNN and WebGPU implementations)
• Offline usage, including interaction with PWAs.
• Common models are lower privacy risks

Note: this is both an AI topic and a Storage topic. Input from both communities would be useful and is encouraged!

There were some related presentations on this topic in the WebML IG.

Session goal

Prioritize issues, discuss highest priority issues, define follow-up actions if possible.

Additional session chairs (Optional)

No response

Who can attend

Anyone may attend (Default)

IRC channel (Optional)

#ai-model

Other sessions where we should avoid scheduling conflicts (Optional)

#23, #36, #39

Instructions for meeting planners (Optional)

Need to avoid scheduling conflicts with sessions related to WoT, Smart Cities, or WNIG.

Agenda for the meeting.

1. Review list of issues and add or refine any if necessary (5m)
2. Prioritize issues, identify shortlist for discussion (10m)
3. Discuss potential solutions to high-priority issues (approx 15m each)
   • Expand explanation of each issue, identify stakeholders
   • Discuss possible resolutions
   • Define followup actions and collaborations

Links to calendar

• 2024-09-25, 13:15 - 14:15

Meeting materials

• Slides
• Minutes
• Repo - Please direct followup there, and to the WebML WG

[tpac-breakout-bot]

Thank you for proposing a session!

You may update the session description as needed and at any time before the meeting, but please keep in mind that tooling relies on issue formatting: follow the instructions and leave all headings and other formatting intact in particular. Bots and W3C meeting organizers may also update the description, to fix formatting issues or add links and other relevant information. Please do not revert these changes. Feel free to use comments to raise questions.

Do not expect formal approval; W3C meeting organizers endeavor to schedule all proposed sessions that are in scope for a breakout. Actual scheduling should take place shortly before the meeting.

[mmccool]

Slides posted here: https://github.com/webmachinelearning/hybrid-ai/blob/main/presentations/WebML%20Discussion%20-%20Hybrid%20AI%20for%20the%20Web%20-%20AI%20Model%20Management%20TPAC%202024%20Breakout.pdf (summary slides presented in WebML meeting on Sept 23 are in same repo).

[mmccool]

Thanks everyone who attended! Note that draft minutes are now linked from the description. The slides have also been posted in the following repo, which I suggest we use for followup discussion (e.g. via issues): https://github.com/webmachinelearning/hybrid-ai/

Although the name of the repo is "hybrid-ai" most of the material there (so far) has been focused on local model management.